Keeping tender documents private when you use AI tools

Don't assume anything is fine to paste. Three things decide it. What's in the text. What the buyer's rules say. What the tool does with your input.

Sort the contents first

A tender pack and your response usually mix several kinds of material:

• Public text the buyer has published for anyone.
• Buyer information that isn't public.
• Your own pricing, methods and strategy.
• Personal details, such as staff CVs, referee names or contact numbers.

Each kind has a different risk. The last one has the clearest official guidance in Australia, so start there.

Personal information

Australia's privacy regulator advises organisations not to enter personal information, especially sensitive information, into public generative AI tools.¹ It also cautions against adopting an AI product merely because it exists.¹

That guidance is written for organisations covered by the Privacy Act.¹ Businesses in the United States, the UK and elsewhere face different laws, so check your own regulator. If a CV or a referee list isn't needed for the task, leave it out.

The buyer's rules

Read the tender for anything on confidentiality, security, copyright or AI. Some buyers state limits plainly. The UK government note for central buyers mentions controls so suppliers don't use confidential tender information as training data. Its example is using confidential tender documents to train a model that writes future bids.²

That note applies to UK central government, and each buyer sets its own terms. Still, it shows what a careful buyer worries about: tender text feeding a model that later helps someone else.

If the tender is silent, ask the buyer in writing, through the channel it names, before you paste.

The tool's own terms

Settings and plans differ a lot, so look at the maker's documents for the exact plan you'd use. Take one maker as an example. Its enterprise privacy page, updated 8 January 2026, says that by default it does not use business data for training. It also says it uses data from versions of its service for individuals.³

Read those two lines together. Your pricing might be treated one way on a business plan and another way on a personal account. The difference is in the plan, not the brand.

Retention is a separate question, and it depends on the product. The maker's page gives two different 30-day periods. For the developer API, the maker may keep inputs and outputs for up to 30 days. Zero data retention can be requested for eligible endpoints only "if you have a qualifying use-case". For workspaces, deleted conversations are removed from its systems within 30 days "unless we are legally required to retain them". On its Business plan the exception is wider. Longer retention can also apply if it is reasonably necessary to protect its services or any third party from harm. Workspace admins control how long data is retained. Check the page for your own plan.³ Other products have their own rules. Inside a team, admins may be able to read and delete what staff type in. So a colleague with admin rights may see what you paste.³

Questions to put to any tool

Ask these before you spend money on it, and write down the date you got each answer:

• Is my input used to train models on this plan, and can I switch that off?
• How long is it kept, and can I delete it?
• Who at the vendor can read it, and under what conditions?
• Who in my own business can see other people's chats?
• Where is the data stored, and does that matter for my buyer?
• Does the contract or data processing agreement cover it?

Maker pages change. Cite the page, note the date, and check again before each big bid.

Paste less

Even with good settings, you can cut risk by pasting less. Use extracts, not whole packs. Swap names for labels. Keep final prices out until the last human step. Then compare the output against the real document yourself.

For what buyers ask suppliers to declare, go to buyer rules on AI use.